Question

我尝试将topk例程与dask隔离开来。它不知何故孤立地死去。

显然，numpy数组而不是dask数组在递归期间传递给x参数。

topk的原始源代码位于：https://github.com/dask/dask/blob/master/dask/array/routines.py

测试程序：

import numpy as np
import dask.array as da
from dask.base import tokenize
from operator import getitem
import dask.sharedict as sharedict
from dask.array.core import Array

def topk(k, x):
    if x.ndim != 1:
        raise ValueError("Topk only works on arrays of one dimension")

    token = tokenize(k, x)
    name = 'chunk.topk-' + token
    dsk = {(name, i): (topk, k, key)
           for i, key in enumerate(x.__dask_keys__())}
    name2 = 'topk-' + token
    dsk[(name2, 0)] = (getitem, (np.sort, (np.concatenate, list(dsk))),
                       slice(-1, -k - 1, -1))
    chunks = ((k,),)

    return Array(sharedict.merge((name2, dsk), x.dask), name2, chunks, dtype=x.dtype)

def main():
    x = np.arange(12)*8
    y = da.from_array(x, 7)
    print(y.topk(2).compute())
    print(topk(2, y).compute())

main()

错误：

  File "test_dask_argtopk.py", line 40, in <module>
    main()
  File "test_dask_argtopk.py", line 38, in main
    print(topk(2, y).compute())
  File "test_dask_argtopk.py", line 27, in topk
    for i, key in enumerate(x.__dask_keys__())}
AttributeError: 'Array' object has no attribute '__dask_keys__'

隔离dask的topk例程

0 个答案: